The Light and Shadow of AI First — The Real Question Is “For Whom”

“AI First failed” and “AI replaces people” are both half-truths. Klarna cut 700 staff with AI and is now rehiring; Microsoft writes 30% of its code with AI and is cutting 6,000 jobs. Same technology, different outcomes. The question is not whether to use AI, but where, and by whom.

1. The reckoning of the “AI First” experiment — the Klarna lesson

In 2024, Klarna CEO Sebastian Siemiatkowski declared with confidence: “AI is doing the work of 700 customer service agents.” A 22% headcount cut followed. Investors cheered, and the tech press anointed Klarna as the model “AI First company.”

A year and a half later, Klarna is hiring people back.

What happened? The AI chatbot handled two-thirds to three-quarters of all customer inquiries, and on paper the numbers looked fine. But the quality of service that customers actually experienced declined. Simple FAQ responses took more than 20 seconds, and past the first exchange “the cracks began to show (PolyAI).” Complex payment disputes, emotional complaints, repeated claims that required contextual understanding — in these domains the AI defaulted to canned “according to our policy…” replies, and the result was a 25% increase in unresolved inquiries. Customers felt they were “fighting with a machine.”

Siemiatkowski himself admitted it. “Cost was the dominant evaluation criterion. As a result, quality went down.” “We pushed AI adoption too aggressively (Bloomberg, Fast Company).” Klarna is now pivoting to an Uber-style flexible workforce model — recruiting students, stay-at-home parents, and rural residents as remote agents — and building a hybrid AI-plus-human model (Fortune).

Klarna is not an outlier. Salesforce followed a similar arc. After cutting 4,000 customer support staff and deploying AI agents, leadership publicly conceded that it had “overestimated AI’s real-world readiness” (CMSWire). The automated systems struggled with nuanced issues, escalations, and long-tail customer problems; service quality fell and complaints rose. Technical limits also surfaced — Salesforce’s field experience revealed that LLMs start ignoring some directives once more than eight are stacked on them.

Gartner’s February 2026 report paints the broader picture. Half of the companies that cut staff because of AI will rehire by 2027. Interestingly, the titles shift. Not “customer service representative” but “solution consultant,” “trusted advisor,” “product specialist.” Only 20% of customer service leaders actually shrank their teams. The majority kept headcount and used AI to serve more customers.

Meanwhile, a Qualtrics survey reports that one in five consumers who used AI customer service got no help at all — a 4x failure rate compared with AI use in other domains. The capital is similarly lopsided. Of the $47 billion poured into AI in the first half of 2025, 89% generated negligible returns.

“AI First” did not fail. Indiscriminate AI First failed. And the essence of that indiscriminateness was shoving AI into work without distinguishing between what AI does well and what it does poorly.

2. And yet — AI is actually working in some places

If you read Klarna and conclude “AI is hype,” you are also half-right. In the same period, AI is delivering tangible results in clearly identifiable places. The difference is what gets handed to AI.

Look at customer service success stories. Freshworks’ Freddy AI automates 53% of retail customer inquiries, cutting first response time from 12 minutes to 12 seconds and resolution time from over an hour to two minutes. GrandStay Hotels saw customer satisfaction rise 22% after AI deployment (Freshworks, Sobot). They have one thing in common: AI is applied only to a narrow, structured slice of inquiries, and complex issues are routed to humans from the start. Companies achieving 210% ROI over three years and payback in under six months all share the same recipe — “clean data, integrated systems, a narrow initial scope, and clear escalation paths.”

The structure of customer service becomes clearer when you look at the data. 60–70% of all inquiries are simple — password resets, order status checks, FAQs. AI resolution rates in this band reach 96–97%, with customer satisfaction equal to or higher than human agents. 61% of consumers actually prefer self-service for simple matters. Meanwhile, 30–40% are complex — disputes, emotional complaints, policy exceptions — and AI fails systematically here. That is because AI “recognizes patterns and predicts the next token,” not “interprets meaning” (HBS Online). The difference between Klarna and Freddy AI is not raw technology but whether this boundary was respected.

The same pattern holds in software engineering. Microsoft has shifted 20–30% of its codebase generation to AI and cut 6,000 jobs. But the story changes when you break the number down. The code AI writes is mostly boilerplate — getters/setters, CRUD operations, simple functions. Analysts peg AI’s coding ability at “roughly a junior developer level.” Of the 2,000 jobs cut in Washington state, more than 40% — 817 software engineers and 373 product managers — were technical roles (Seattle Times), with the official goal being “to flatten management layers.” Senior engineers doing architectural design, system design, and complex debugging remain in place.

Medicine is the most dramatic and the most subtle case. Automation of medical transcription has accelerated, and the BLS projects employment in this occupation to decline 5% from 2023 to 2033 (BLS). But “99% automation” is half-true. AI transcription produces errors that are qualitatively different from human ones. For example, it will confidently transcribe “digoxin 0.25mg” as “digoxin 2.5mg” — a 10x dosing error (DeepCura). As a result, medical transcriptionists have not “disappeared” — their role has shifted from “people who type” to “people who validate AI output.” The headcount shrinks, but the skills required of those who remain rise.

Radiology shows the same shape. The FDA has approved 873 radiology AI algorithms, and 85% of radiologists say AI will improve patient care (RSNA). The early prediction that “AI will replace radiologists” missed. What actually happened is “radiologists who use AI replace radiologists who don’t.” In medical coding, AI cut errors 40% and raised productivity 33% (PMC), but human oversight remains essential.

Lay the three industries side by side and the pattern emerges.

Industry	Tasks AI replaced	Tasks AI did not replace
Customer service	Simple inquiries (FAQ, order status, password)	Dispute resolution, emotional handling, policy exceptions
Software	Boilerplate code, simple function generation	Architectural design, complex debugging, security design
Medicine	Document transcription, basic coding	Clinical judgment, patient-facing work, AI output validation

The common denominator: AI is faster and cheaper than humans on tasks where inputs are structured, outputs are predetermined, and the scope of judgment is narrow. On tasks that require contextual interpretation, exception handling, emotional interaction, and system-level reasoning, AI fails systematically. This is not an “AI capability problem” — it is a structural property of the technology.

3. Not “jobs” but “tasks” — and juniors are at risk

Push one level deeper. Behind the surface phenomenon of “some companies cut, others hire” is a more fundamental pattern that researchers have captured.

MIT economist David Autor updated his 2003 “routine/non-routine” framework for the AI era. The core distinction (Brookings, NBER):

Automation tool: Removes the need for expertise. The person who did the work is no longer needed.
Collaboration tool: Amplifies expertise. People who already know more can do more.

The same AI technology, inside the same job, can act as an automation tool on one task and a collaboration tool on another. In medical transcription, “dictation” was automated, but “validation” emerged as a new collaboration task. In coding, “boilerplate generation” was automated, but “reviewing AI-written code” became more important as a collaboration task.

Here is what makes this wave decisively different from previous ones. Stanford economist Erik Brynjolfsson calls it the “Turing Trap” (Stanford Digital Economy Lab):

“AI that mimics humans (automation) lowers wages. AI that enables humans to do what they couldn’t before (amplification) raises them.”

Past automation mostly replaced mid-skill routine work — factory assembly, clerical support, data entry. LLM-based AI reaches into high-skill cognitive labor. This is an inversion of the prior automation wave. Tasks once considered “safe” in high-education, high-income occupations — legal document analysis, code generation, diagnostic support — are now the automation target.

The data, though, defies intuition. According to the Dallas Fed’s February 2026 analysis, wages actually rose in occupations with high experience premiums — lawyers, insurance underwriters, credit analysts — after AI adoption. AI cuts the three-hour case-law search a lawyer does down to ten minutes. But interpreting those cases, designing strategy, and persuading clients still belong to the lawyer. AI worked as a “collaboration tool,” not an “automation tool.”

In occupations with low experience premiums — ticketing agents, fast-food workers, data entry clerks — employment fell and wages stagnated. When a 10-year veteran and a one-year hire process work at roughly the same speed, there is no reason to keep paying for human labor once AI can do it. Wages in computer systems design rose 16.7% while the national average grew just 7.5%.

And here is the most painful data point. A 2025 Harvard/SSRN study analyzed 285,000 U.S. firms and 62 million workers. After a firm adopts generative AI, junior hiring drops 9–10% within six quarters. Senior employment barely moves. Crucially, this is not “layoffs” — it is “hiring freezes.” Open slots simply do not get filled.

Stanford’s data is more granular. Employment of software developers aged 22–25 is down nearly 20% from the 2022 peak (Understanding AI). Entry-level hiring at the top 15 tech companies fell 25% in 2023–2024. In a survey of 500 tech leaders, 72% said they planned to cut entry-level developer hiring, and 64% said they would “invest in AI tools instead” (CIO).

This produces a structural problem. Junior developers learn by writing boilerplate. They internalize databases by implementing simple CRUD; they absorb code structure by writing unit tests. If AI replaces this “apprenticeship,” then five to seven years out, the pipeline breaks — there is no one to become senior. One engineer at Companion Group reported that after heavy AI tool use, “work I used to do instinctively became manual and tedious” (Stack Overflow). Developer positivity toward AI fell from above 70% in 2023–24 to 60% in 2025.

The conclusion all this data points to is the same. AI does not replace “jobs” wholesale — it selectively replaces “tasks.” And the tasks it replaces are predominantly structured, repetitive ones. Whether you make this distinction precisely when designing your AI deployment is what separates success from failure — that is exactly what divides Klarna from Freshworks.

4. The trap of “every employee an AI builder”

In this context, we need to talk about a recently noticed startup.

Gumloop raised $50 million in a Series B in March 2026. Branding itself as a no-code AI agent builder, it pitches a vision of “a world where every employee can build AI agents.” Drag-and-drop workflow design, AI agents deployable without writing a line of code.

It is an attractive vision. If non-developers can put AI to work, organizational productivity should rise. That is true — in theory.

The data tells a different story. The RAND Corporation’s 2025 report puts the failure rate of AI projects at 80.3%. Narrowed to generative AI pilots, 95% fail to move past proof of concept into scaled deployment. Of the $684 billion poured into AI globally, one analysis estimates more than 80% has fallen short of expectations (Pertama Partners).

Why? Not because tooling is scarce. Because the ability to define problems is scarce.

What a no-code AI builder solves is the “implementation barrier.” You can build an agent without knowing how to code. But “which part of this workflow should be automated?” “How will input data quality be guaranteed?” “How will we detect and recover when the agent makes a wrong call?” — answering these is a job for human capability, not for tools. No matter how easy no-code gets, if you don’t know what to build, easier tools just let you build the wrong thing faster.

And there is data suggesting that the very sense that “AI lets me build fast” may be an illusion. METR’s 2025 randomized controlled trial produced a startling result. Sixteen experienced open-source developers (average repo maintainers with over 22,000 stars) were given AI tools and real tasks. The AI-using group was actually 19% slower. And yet they believed they were 20% faster. That perception did not change even after the experiment ended. The causes: imperfect prompting, wrestling with the AI interface, mismatches between high quality standards and AI suggestions, and the cognitive cost of “experimenting” with AI.

Code quality data is even more direct. CodeRabbit’s 2025 analysis compared 470 GitHub PRs (320 co-authored with AI, 150 human-only). Code that involved AI had 10.83 issues per PR vs. 6.45 for human-only — roughly 1.7x more. And not just more bugs — different kinds:

Logic/correctness errors: +75%
Security vulnerabilities: 1.5–2x (poor credential handling, insecure object references)
Code readability problems: 3x
Performance inefficiencies (excess I/O): nearly 8x
Concurrency control errors: 2.29x

GitClear’s five-year analysis of 211 million lines of code (including Google, Microsoft, Meta) shows codebase-level shifts. Copy/paste code rose from 8.3% in 2020 to 12.3% in 2024 — a 48% increase. Refactored code fell from 24.1% to 9.5% — more than halved. Code churn (new code modified or reverted within two weeks) rose from 5.5% to 7.9%. AI generates code faster, but the time spent fixing that code is rising too.

So who are the real beneficiaries of no-code AI builders? People who understand their workflows, can judge where automation belongs, and can validate outputs. For them, no-code becomes “a tool that shortens implementation time.” Hand the same tool to someone without that judgment, and you get fast wrong answers.

Making tools easier and designing where to apply them are entirely different problems. What decides the success or failure of AI adoption is not tool capability but the expertise to design the precise application points.

5. History repeats — but the speed is different this time

It would be reasonable to feel uneasy. Junior hiring is shrinking, AI failure rates are 80%, code quality is falling. The press is good at cherry-picking these numbers to manufacture fear. But this is not the first time technology has reshaped labor markets. Historical patterns sharpen the picture of what is happening now.

ATMs and bank tellers (1970s–). When ATMs arrived, every expert predicted the end of bank tellers. Tellers per branch did fall, from 20 in 1988 to 13 in 2004. But as branch operating costs dropped, banks opened 43% more branches in urban areas. Net teller employment grew faster than the labor force average (AEI). The teller role shifted from “counting cash” to “relationship management and product advising.”

Spreadsheets and accountants (1980s–). With VisiCalc (1979), Lotus 1-2-3, and Excel, manual calculations that took 20 hours shrank to 15 minutes. The result: bookkeepers fell by 400,000, while licensed accountants grew by 600,000 (NPR). As calculation got cheap, firms began demanding scenario analysis and strategic modeling they had never attempted before, and demand exploded for people who could “interpret” numbers rather than “produce” them.

The same four-stage pattern repeats in both cases:

Routine execution tasks get automated (cash counting, manual calculation, manual drafting)
The cost of that function falls, demand itself grows (more branches, more scenario analysis)
The value of remaining human tasks rises (relationship management, strategic interpretation, design judgment)
Net employment in the broader field is sustained or grows

The same pattern is starting to appear in AI. As code generation gets cheap, total software project volume grows, and the value of architecture and system design rises. As transcription cost disappears in medicine, clinical record volume explodes and demand grows for the capability to validate and analyze it.

But there is one decisive difference: speed.

The ATM transition took 40 years. Over that span, tellers had time to retrain into new roles. With spreadsheets, bookkeepers had time to study, certify, and become accountants. The AI transition is happening in three years. Junior developer hiring dropped 20% in two years. Klarna cut 700 staff and started rehiring within 18 months. There is an overwhelming shortage of time for retraining and adaptation.

This is the real lesson the historical pattern offers. Over the long run, AI will create new roles. But in the short run, when the speed of transition outpaces the speed of adaptation, people get hurt in between. The long-term forecast that “things will be fine eventually” may be right. But the steepness of getting to that “eventually” — far sharper this time — is the distinctive risk.

6. The right answer is “AI Effective” — with conditions

So what is the right approach?

A concept emerging in industry: “AI Effective” — a design philosophy of applying AI not to everything, but only where AI is effective.

AI First means: “Apply AI to every process first, and pull it out of places where it doesn’t work later.” Aggressive and fast. But as Klarna showed, the cost of discovering “where it doesn’t work” after the fact is high. Customer churn, quality decline, rehiring expense. Customers and employees pay the price of the experiment.

AI Effective is the opposite. “Identify where AI is effective first, and apply it only there.” Freshworks routing only 53% of retail inquiries to AI and leaving the rest to humans is this approach. Microsoft directing AI at boilerplate while leaving architecture to seniors is the same. So is medicine handing transcription to AI while routing validation to experienced transcriptionists.

AI Effective is correct. I agree. But there is a precondition.

To identify “where AI is effective,” you first need a deep understanding of the work itself. Knowing where the line falls between the 60% of simple customer service inquiries and the 40% of complex ones requires customer service operations experience. Knowing where boilerplate ends and core logic begins requires understanding of software architecture. Catching the 1% of confidently-wrong outputs in medical transcription requires medical domain knowledge.

Judging the “Effective” in AI Effective is itself the realm of expertise. Tools do not solve this. No-code does not solve this. To avoid Brynjolfsson’s “Turing Trap” — that is, to design AI as a tool that “amplifies humans” rather than “replaces” them — you need the eye to distinguish what is replaceable from what is amplifiable. That eye is born at the intersection of field experience and technical understanding.

So what gains value in the AI era is, paradoxically, human expertise. The capability to decompose the tasks in your own work and design the precise boundary between the 60% AI takes and the 40% humans keep. There is an enormous cost difference between knowing this from the start and learning it through 18 months of Klarna-style trial and error. In AI adoption, choosing a partner capable of “right design” decides the outcome.

7. Conclusion — what makes the same AI produce different results is “design”

To summarize.

“AI is taking jobs” — half-true. What AI replaces is not “jobs” but “tasks.” And those tasks are predominantly structured, repetitive ones. Dallas Fed data, Harvard research, Stanford data all point the same direction — AI is an amplifier for people with expertise and a substitute for routine work.

“AI First failed” — half-true. What failed was indiscriminate AI First that did not separate the 60% from the 40% of customer service, did not separate boilerplate from architecture, did not separate transcription from validation. Companies that applied AI precisely at the task level are delivering 210% ROI and 50% cost reduction.

“Every employee should become an AI builder” — the direction is right, but the premise is missing. As the METR experiment showed, simply using AI tools does not make you faster. Easier tools do not deliver the capability to design “where to apply them.” The difference between Freshworks and Klarna was not a technology gap but a design gap.

“Things will work out in the long run” — by historical pattern, probably yes. After ATMs, bank employment grew. After spreadsheets, accountant employment grew. But a 40-year transition is happening in three years. So what matters right now is design that can withstand the speed of transition.

In the end, the most expensive mistake in AI adoption is not “failing to adopt AI.” The most expensive mistake is adopting AI without designing where to apply it. Klarna paid for that lesson with 18 months of customer churn and rehiring costs. Salesforce had to publicly admit it had “overestimated AI’s real-world readiness” after cutting 4,000 jobs.

In the same period, Freshworks handed 53% of inquiries to AI and delivered 210% ROI. Microsoft pointed AI at boilerplate and lifted senior engineer productivity. Medicine handed transcription to AI but kept validation with experienced specialists.

What made the difference was not technology but the expertise to design precisely where AI takes over and where humans take it from there. Knowing the boundary between AI’s 60% and the human 40%. Building the capability to validate AI-written code’s 2x security vulnerabilities and 2.3x concurrency errors. Securing the domain knowledge to catch the 1% AI transcription is “confidently wrong” about.

The most valuable capability in the AI era is not AI itself. It is the design capability to place AI in the right places. And that design comes only from places that understand both technology and business.